fix(gateway): stop synthetic goal/heartbeat turns from stomping channel_pin by chelsealong · Pull Request #126155 · NousResearch/hermes-agent

chelsealong · 2026-09-28T07:55:04Z

Root cause

Gateway turns injected from synthetic events (goal continuation, heartbeat
due-prompts, /goal resume) are built as non-internal MessageEvents
(they must still pass user authorization) with channel_prompt=None.

_pinned_channel_inputs() (gateway/run_agent_cache.py) only reused the
session's pinned channel inputs for internal events — every non-internal
event, synthetic or not, recorded (channel_prompt, source.parent_chat_id)
as the new channel_pin. Since these synthetic events never carry their own
channel_prompt, that recorded (None, parent_chat_id), dropping the
channel prompt / channel_overrides prompt for that turn and then flipping
it back on the next human turn — the same A→B→A pattern as #125763 and
#124179 (shorter ephemeral system prompt, changed _agent_config_signature,
cached-agent rebuild, prefix-cache miss).

Producers, all confirmed on current main:

gateway/run_goals.py::_synthetic_prompt_event — used by goal
continuation (_post_turn_goal_continuation) and the heartbeat due-prompt
poller (_heartbeat_poll_watch)
gateway/slash_commands_goals.py::_enqueue_goal_turn — /goal resume
continuation (channel_prompt=event.channel_prompt if kickoff else None)

Fix

internal was overloaded as the proxy for "reuse the channel pin", but it
also gates authorization bypass (and touch_activity) — heartbeat and goal
events deliberately stay non-internal so those checks still apply. Added a
separate MessageEvent.inherit_channel_pin flag and set it on the three
synthetic producers above (the /goal resume case only, not the kickoff
case, since kickoff carries its own real channel_prompt). Both call sites
of _pinned_channel_inputs now pass reuse_pin=event.internal or event.inherit_channel_pin.

Test plan

Added test_synthetic_non_internal_event_reuses_channel_pin to
tests/gateway/test_internal_event_pin_wiring.py, which drives the real
_handle_message_with_agent handler through human → synthetic (non-internal,
inherit_channel_pin=True, channel_prompt=None) → human turns on one
Discord thread session with a channel_overrides parent persona configured,
and asserts the resolved channel_prompt passed to _run_agent is
"Channel hint." on all three turns (never flips to None), and that the
combined ephemeral prompt keeps both "Channel hint." and "Parent persona." throughout.

Confirmed the test fails without the fix (reverted the 5 source files,
left the test in place):

FAILED tests/gateway/test_internal_event_pin_wiring.py::test_synthetic_non_internal_event_reuses_channel_pin
E       TypeError: MessageEvent.__init__() got an unexpected keyword argument 'inherit_channel_pin'

With the fix restored, the new test and the rest of the suite pass:

$ scripts/run_tests.sh tests/gateway/test_internal_event_pin_wiring.py
=== Summary: 1 files, 3 tests passed, 0 failed (100% complete) in 2.3s (4 workers) ===

Also ran the full set of gateway tests that exercise MessageEvent,
channel_prompt/channel_pin, goal continuation, or heartbeats (to catch
any test double missing the new field), plus tests/hermes_cli/test_goal_dispatch.py:

$ scripts/run_tests.sh <64 gateway test files matching heartbeat/goal_continuation/GatewayGoalsMixin>
=== Summary: 64 files, 1224 tests passed, 0 failed, 7 skipped (100% complete) in 70.1s (4 workers) ===

$ scripts/run_tests.sh tests/gateway/test_cached_agent_history_guard.py tests/gateway/test_discord_channel_prompts.py \
    tests/gateway/test_discord_triggering_note_persistence.py tests/gateway/test_feishu_channel_prompts.py \
    tests/gateway/test_gateway_silence_tokens.py tests/gateway/test_queue_command.py \
    tests/gateway/test_queued_final_ledger.py tests/gateway/test_telegram_group_gating.py \
    tests/gateway/test_telegram_mention_context.py tests/gateway/test_voice_command.py \
    tests/gateway/test_internal_event_pin_wiring.py tests/hermes_cli/test_goal_dispatch.py
=== Summary: 12 files, 200 tests passed, 0 failed, 4 skipped (100% complete) in 21.1s (4 workers) ===

This first pass caught two existing tests (test_queued_final_ledger.py,
test_discord_triggering_note_persistence.py) that build pending_event
as a bare SimpleNamespace without the new field; switched both
_pinned_channel_inputs call sites to getattr(event, "inherit_channel_pin", False) (matching the existing defensive getattr(event, "internal", False)
pattern used elsewhere) so the new field is optional for callers/mocks, and
reran — all green.

ruff check passes on all changed files.

AI assistance disclosure

This PR was written by an autonomous AI coding agent (Claude, Anthropic)
operating against this repository's own issue tracker and test suite, with
no human-authored code changes.

🤖 Generated with Claude Code

…el_pin Goal continuation, heartbeat due-prompts and /goal resume events are built non-internal (they must still pass authorization) with channel_prompt=None. _pinned_channel_inputs() only reused the pin for internal events, so these synthetic turns recorded (None, parent_chat_id) as the new channel_pin, dropping the channel prompt and channel_overrides for that turn and flipping them back on the next human turn (NousResearch#125763/NousResearch#124179-style A->B->A: shorter ephemeral prompt, changed _agent_config_signature, cached-agent rebuild, prefix-cache miss). Add MessageEvent.inherit_channel_pin, independent of `internal` (which also gates authorization bypass and touch_activity), and set it on the three producers that carry no channel inputs of their own. _pinned_channel_inputs now reuses the pin for internal OR pin-inheriting events. Fixes NousResearch#126109

ehz0ah

I reviewed this at abe539ac. The focused suites pass as submitted: 72 tests passed across the pin-wiring and goal-dispatch files. The channel-pin change is directionally correct, but the full prompt still changes on the synthetic turn. A one-line byte-equality assertion in the new real-handler regression fails on this head. Details are inline. This head also conflicts with current main in the pin-wiring test and has no hosted checks.

ehz0ah · 2026-09-28T08:04:52Z

            _turn_channel_prompt, _turn_source = self._pinned_channel_inputs(
-                session_key, event.channel_prompt, source, internal=event.internal,
+                session_key, event.channel_prompt, source,
+                reuse_pin=event.internal or getattr(event, "inherit_channel_pin", False),


Blocking: this reuses only the channel pin. Earlier in the same handler, _hmwa_prepare_turn() still calls _pinned_session_context_prompt(..., internal=event.internal). _synthetic_prompt_event() removes message_id but keeps the event non-internal, and Discord includes message-ID presence in _ephemeral_change_key(). The synthetic turn therefore rebuilds a shorter session-context block, and the next human turn changes it back. I added only assert eph[0] == eph[1] == eph[2] to this PR’s new real-handler test; it failed with 2 passing tests and 1 failure because the middle prompt lost the “Triggering message” line. Please apply the same inheritance signal to the session-context pin and keep that full byte-equality assertion. That closes the A→B→A invariant rather than only preserving channel_prompt.

ehz0ah · 2026-09-28T08:04:52Z

+    # heartbeat due-prompts, /goal resume) so the turn reuses the session's pinned channel
+    # prompt/parent instead of overwriting it with this event's own None (#126109). Independent
+    # of ``internal``: these events still go through user authorization.
+    inherit_channel_pin: bool = False


Please append this field after the existing init fields instead of inserting it before metadata. MessageEvent is the normalized adapter boundary, and its own field-order comment requires new fields to be appended for positional compatibility. This placement shifts metadata, timestamp, allow_gateway_control, and reply_expected for external adapters or plugins that still construct events positionally.

teknium1

PR Review — #126155 (head `abe539acdd303cdff13bddc4bd8a739dfa03bf58`)

Verdict: request changes — the diagnosis is right and the channel_pin half is fixed, but the fix is partial and the head conflicts with main. Merge decision is Teknium's.

Context. Synthetic goal-continuation / heartbeat / /goal resume turns are non-internal and carry channel_prompt=None; _pinned_channel_inputs treated every non-internal event as a human turn and overwrote the pin with (None, parent) (#126109). Your inherit_channel_pin flag on the three producers fixes that half; verified by driving the real handler on your tree merged onto origin/main @ 188a1a5: channel_prompt == ['Channel hint.'] * 3 (45 tests passed across your test file + goal/heartbeat/queue siblings).

Blocking — the session-context pin still flips (ehz0ah's inline finding, confirmed live). run_turn.py::_hmwa_prepare_turn still calls _pinned_session_context_prompt(..., internal=event.internal). _synthetic_prompt_event strips source.message_id, and on Discord _ephemeral_change_key keys on message-id presence, so the synthetic turn re-renders the session-context block and the next human turn re-renders it back. Same probe on your tree: context bytes 1238 → 1122 → 1238 while the channel prompt stays stable; the whole point of the pin is byte-stable system bytes, so half the A→B→A (and the agent rebuild) survives. The issue triage lists both _pinned_session_context_prompt and _pinned_channel_inputs as the two seams; feed the same internal or inherit_channel_pin decision to both (#126167 does this and its byte-equality test passes; your test should assert ephemeral[0] == ephemeral[1] == ephemeral[2] too, which is what makes it red on main for the right reason — the current red-on-base is a TypeError on the missing kwarg, not the symptom).

Also needed:

gateway/platforms/event.py: MessageEvent's own rule (comment above allow_gateway_control) is that new fields append after reply_expected for positional compat; inherit_channel_pin is inserted before metadata.
Rebase: tests/gateway/test_internal_event_pin_wiring.py conflicts with main (#126092 added tests in the same slot — keep both). GitHub dispatches zero workflows on a conflicting PR, which is why the branch shows no checks.
Minor: the getattr(event, "inherit_channel_pin", False) at the two call sites is only there for SimpleNamespace doubles in two tests; fine, but if you take the _event_preserves_prompt_pins-style helper it lives in one place.

Overlap: #126167 (filed six minutes after this one) covers both pins and passes the full byte-equality probe; you found the channel half first and should be credited whichever lands. If you update this PR to cover the context pin as above, the two are equivalent and the choice is the maintainer's.

teknium1 · 2026-09-28T08:45:14Z

+    # heartbeat due-prompts, /goal resume) so the turn reuses the session's pinned channel
+    # prompt/parent instead of overwriting it with this event's own None (#126109). Independent
+    # of ``internal``: these events still go through user authorization.
+    inherit_channel_pin: bool = False


MessageEvent's own rule (comment above allow_gateway_control): new fields append after reply_expected for positional compat. Move inherit_channel_pin to the end of the init fields.

teknium1 · 2026-09-28T08:45:14Z

        if not session_key:
            return channel_prompt, source
-        if not internal:
+        if not reuse_pin:


This fixes the channel half only. _hmwa_prepare_turn still calls _pinned_session_context_prompt(..., internal=event.internal), and because _synthetic_prompt_event strips message_id (a _ephemeral_change_key input on Discord), the session-context pin still flips. Live on this head merged onto main: context bytes 1238 → 1122 → 1238 across human → synthetic → human while channel_prompt stays 'Channel hint.'. Pass the same internal or inherit_channel_pin decision to _pinned_session_context_prompt, and assert full ephemeral byte-equality in the test.

Feed the same internal or inherit_channel_pin decision into _pinned_session_context_prompt as _pinned_channel_inputs, so a synthetic goal/heartbeat/resume turn (which strips message_id, a _ephemeral_change_key input) doesn't flip the session-context block and back on the next human turn. Also moves inherit_channel_pin to append after reply_expected for MessageEvent positional compat.

…c-channel-pin # Conflicts: # tests/gateway/test_internal_event_pin_wiring.py

chelsealong · 2026-09-28T10:11:45Z

Addressed both blocking points at 10df6fd2b5:

run_turn.py::_hmwa_prepare_turn now calls _pinned_session_context_prompt(..., internal=event.internal or getattr(event, "inherit_channel_pin", False)), so the session-context pin reuses on the same decision as the channel pin. Live probe (revert just run_turn.py/run_agent_cache.py, keep the new test): the byte-equality assertion in the merged test_synthetic_non_internal_event_reuses_channel_pin fails with "synthetic turn toggled the ephemeral prompt bytes" (the "Triggering message" line drops on the synthetic turn), confirming red-on-base for the actual A→B→A. With the fix restored it's green.
MessageEvent.inherit_channel_pin moved to append after reply_expected, per the field-order rule.
Merged upstream/main (not rebased, to avoid a force-push) to pick up Leftover steer / interrupt follow-ups keep the channel prompt (no agent rebuild) (#125763, salvage #125783) #126092's new tests in the same file; both sets of tests plus the updated regression test all pass together (10/10 in the file).

scripts/run_tests.sh tests/gateway/test_internal_event_pin_wiring.py → 10 passed. Also reran the full gateway/goal/heartbeat/channel-prompt sweep from the PR description (13 files, 239 tests) plus ruff check on all changed files — all green.

ehz0ah

Reviewed exact head 10df6fd2b5361e4a8db144c836b6b6950ce76b9f.

The new commits resolve both blockers from my earlier review. Synthetic goal, heartbeat, and resume turns now reuse the session-context pin as well as the channel pin, so dropping message_id no longer changes the full ephemeral prompt. The new MessageEvent field is also appended after the existing init fields, which preserves positional compatibility.

I traced the producer and consumer paths. The affected events remain non-internal, so authorization, emergency-stop handling, and session activity updates are unchanged. The effective-prompt regression now checks byte-identical prompts across a human turn, a synthetic turn, and the next human turn.

Local verification passed: 108 tests across the pin-wiring, goal continuation and resume, heartbeat, queue, and goal-dispatch suites. GitHub currently reports no hosted checks.

One non-blocking coverage improvement would be to assert inherit_channel_pin directly in the heartbeat, goal-continuation, and resume producer tests. The current end-to-end regression constructs the flag explicitly, so those assertions would catch a future producer-wiring regression closer to its source.

I found no further code-level blocker on this head. GitHub currently reports a merge conflict with main, so the branch still needs to be updated before it can merge.

chelsealong · 2026-09-30T09:12:30Z

Merged current upstream/main to clear the conflict (kept both sides in run_agent_cache.py docstrings and in test_internal_event_pin_wiring.py); scripts/run_tests.sh tests/gateway/test_internal_event_pin_wiring.py → 13 passed. No other code changes; the optional producer-level inherit_channel_pin assertions are not included, to keep the PR small.

teknium1 · 2026-10-16T22:06:05Z

Salvaged into #139865, merged as 8eecd1c, with your commit and authorship kept. The inherit_channel_pin field on MessageEvent is the mechanism that shipped. Thank you @chelsealong.

JoaoMarcos44 mentioned this pull request Sep 28, 2026

fix(gateway): preserve prompt pins and pending turn ownership #126167

Open

ehz0ah reviewed Sep 28, 2026

View reviewed changes

teknium1 mentioned this pull request Sep 28, 2026

fix(gateway): persist internal prompt pins across restart (salvage #125832) #126137

Merged

teknium1 requested changes Sep 28, 2026

View reviewed changes

chelsealong added 2 commits September 28, 2026 10:02

Merge remote-tracking branch 'upstream/main' into fix-126109-syntheti…

10df6fd

…c-channel-pin # Conflicts: # tests/gateway/test_internal_event_pin_wiring.py

ehz0ah reviewed Sep 28, 2026

View reviewed changes

EdderTalmor mentioned this pull request Sep 29, 2026

fix(gateway): preserve prompt pins across synthetic goal/heartbeat/resume turns #127208

Closed

creditlabsonarc-rgb added a commit to creditlabsonarc-rgb/hermes-agent that referenced this pull request Sep 29, 2026

chore: claim P0 NousResearch#126155 [WIP]

cfe27d6

creditlabsonarc-rgb mentioned this pull request Sep 29, 2026

fix(P0): fix(gateway): stop synthetic goal/heartbeat turns from stomping channe #127487

Draft

jonpol01 mentioned this pull request Oct 1, 2026

[Bug]: a session opened by /goal never loads the channel's bound skill (kickoff drops auto_skill; a non-generating command spends first-turn loading) #130316

Open

2 of 3 tasks

teknium1 mentioned this pull request Oct 4, 2026

fix(gateway): never downgrade channel pin to None on synthetic turns (fixes #126109) #132681

Closed

teknium1 mentioned this pull request Oct 16, 2026

fix(gateway): synthetic goal/heartbeat turns and post-eviction wakes keep the system prompt byte-stable (#126109, #131031, salvage #131032 #126155) #139865

Merged

teknium1 closed this Oct 16, 2026

Navigation Menu

Repository navigation

fix(gateway): stop synthetic goal/heartbeat turns from stomping channel_pin - #126155

fix(gateway): stop synthetic goal/heartbeat turns from stomping channel_pin#126155
chelsealong wants to merge 3 commits into
NousResearch:mainfrom
chelsealong:fix-126109-synthetic-channel-pin

chelsealong commented Sep 28, 2026

Uh oh!

ehz0ah left a comment

Uh oh!

ehz0ah Sep 28, 2026

Uh oh!

ehz0ah Sep 28, 2026

Uh oh!

teknium1 left a comment

Uh oh!

teknium1 Sep 28, 2026

Uh oh!

teknium1 Sep 28, 2026

Uh oh!

chelsealong commented Sep 28, 2026

Uh oh!

ehz0ah left a comment

Uh oh!

chelsealong commented Sep 30, 2026

Uh oh!

teknium1 commented Oct 16, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

4 participants

Conversation

chelsealong commented Sep 28, 2026

Root cause

Fix

Test plan

AI assistance disclosure

Uh oh!

ehz0ah left a comment

Choose a reason for hiding this comment

Uh oh!

ehz0ah Sep 28, 2026

Choose a reason for hiding this comment

Uh oh!

ehz0ah Sep 28, 2026

Choose a reason for hiding this comment

Uh oh!

teknium1 left a comment

Choose a reason for hiding this comment

PR Review — #126155 (head abe539acdd303cdff13bddc4bd8a739dfa03bf58)

Uh oh!

teknium1 Sep 28, 2026

Choose a reason for hiding this comment

Uh oh!

teknium1 Sep 28, 2026

Choose a reason for hiding this comment

Uh oh!

chelsealong commented Sep 28, 2026

Uh oh!

ehz0ah left a comment

Choose a reason for hiding this comment

Uh oh!

chelsealong commented Sep 30, 2026

Uh oh!

teknium1 commented Oct 16, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

4 participants

PR Review — #126155 (head `abe539acdd303cdff13bddc4bd8a739dfa03bf58`)